Artificial intelligence is expanding what researchers can do computationally in antibody discovery. Models can generate antibody sequences, predict three-dimensional structures, identify potential binding interfaces, and estimate properties relevant to affinity and developability. In some cases, these approaches allow large numbers of candidate designs to be prioritized computationally before experimental work begins.

Yet a designed sequence remains a hypothesis about a molecule.

Researchers still need to determine whether that sequence can be expressed, whether the resulting antibody binds its intended target, how that interaction behaves, and whether the molecule has biophysical properties compatible with further development.

Antibody data extend far beyond sequence

Large repertoire databases contain enormous numbers of variable-region sequences, providing a rich foundation for antibody language models and other sequence-based approaches.

A recent review illustrates the difference in scale between these resources and other datasets used in antibody discovery. Some antibody language models have been trained on hundreds of millions of sequences, whereas structure-prediction and antibody-antigen interaction models have been developed using datasets containing thousands of antibody structures or complexes from curated resources such as SAbDab.

A sequence describes the amino acid composition of an antibody. Structural information describes how those amino acids may be arranged in three-dimensional space. Interaction data provide information about how an antibody recognizes its antigen. Experimental characterization adds another layer by measuring properties such as expression, binding kinetics, thermostability, polyreactivity and self-interaction.

Together, these measurements create an informative experimental profile.

Predicted structure is not the same as measured behavior

Deep-learning methods have substantially improved protein and antibody structure prediction. Antibody-specific models can now rapidly generate structural representations from sequence, making it possible to evaluate many more candidates computationally.

But prediction accuracy is not uniform across an antibody.

The complementarity-determining regions (CDRs), particularly CDR H3, are important for antigen recognition and are also among the most structurally variable parts of an antibody. CDR H3 modeling remains challenging even as other aspects of antibody structure prediction have improved.

Antibody-antigen interactions introduce additional challenges. Binding depends on the structures of the antibody and antigen, but also on the molecular context of the interaction. Computational docking can be strongly influenced by the availability of epitope or paratope information, and in some cases accurate prediction may be limited by the absence of experimentally determined antibody-antigen complex structures.

For this reason, computational predictions and experimental measurements provide complementary rather than interchangeable information.

The question for experimental characterization becomes: which measurements provide the most useful information about each designed molecule?

Expression provides an early experimental readout

Expression data provide one of the earliest links between a computational design and its experimental behavior. A sequence may appear plausible computationally but still express at a level that makes it less attractive for further development.

For large candidate panels, expression measurements can also reveal variation, so it can be considered an experimental attribute associated with each antibody design.

When expression is integrated with downstream characterization, each candidate begins to accumulate a set of comparable measurements: sequence, expression level, binding characteristics and biophysical properties.

Rapid expression can make this characterization available sooner. Biointron's RushData platform, for example, incorporates a 1-day CHO expression workflow upstream of subsequent antibody assays. The purpose is not simply to produce antibody rapidly, but to move designed sequences into experimental characterization and structured data generation within days.

Binding requires experimental measurement

Computational models can predict antibody-antigen interactions and prioritize mutations for affinity optimization, but these outputs remain predictions. Antibody binding depends on molecular features that are difficult to capture completely in silico. For example, molecular water, hydrogen-bonding networks and environmental conditions can influence antibody-antigen interactions and may be overlooked in computational mutation-prediction approaches.

Experimental assays therefore remain necessary to determine how a candidate actually binds. Biolayer interferometry (BLI) and surface plasmon resonance (SPR) can characterize binding affinity and, depending on the assay design and analysis, association and dissociation kinetics. The equilibrium dissociation constant (K_D) describes affinity, while (k_on) and (k_off) describe the rates of complex formation and dissociation.

Linking these measurements to individual antibody sequences provides experimentally determined binding data rather than a predicted score alone. RushData can incorporate BLI- or SPR-based binding characterization alongside expression and other assay results.

Binding is only one dimension of antibody quality

Besides affinity, candidates can also differ in thermal stability, self-interaction, and nonspecific or polyreactive binding.

These properties are not necessarily independent, as computationally identified aggregation-prone regions can overlap with CDRs, raising the possibility that sequence changes affecting aggregation propensity could also influence antigen recognition.

This makes multi-parameter experimental characterization useful during candidate comparison. Differential scanning fluorimetry (DSF) provides measures of thermal stability such as (T_m), while AC-SINS assesses antibody self-interaction. Polyspecificity-reagent assays provide information about polyreactive or nonspecific binding.

RushData's developability workflow includes DSF (T_m) and (T_onset), PSR-BVP for polyreactivity assessment, and AC-SINS for self-interaction. These measurements can be evaluated together with expression and binding data for the same candidate panel.

From assay results to multidimensional datasets

A candidate profile can include:

• expression and concentration
• binding affinity and kinetic parameters
• analytical characterization
• thermal stability
• polyreactivity
• self-interaction

When these measurements are generated under consistent conditions across a candidate panel, researchers can compare antibodies across several properties.

This is also relevant to machine learning, since the availability of high-quality training data is an ongoing limitation, and greater standardization and availability of experimental data could improve AI-based antibody design.

Generating experimental profiles at scale

As computational approaches increase the number of antibody sequences that can be proposed, experimental workflows must characterize larger candidate panels without reducing each molecule to a single readout.

RushData is designed for this purpose. The platform combines high-throughput, 1-day CHO expression with binding and developability assays and can process more than 3,000 molecules per batch. Results are returned as structured datasets intended for candidate comparison and computational workflows.

‍

About Biointron

Founded in 2012 and certified to ISO 9001:2015, Biointron is a CRO specializing in antibody discovery, expression, and optimization services for biotech and pharmaceutical companies. From gene sequence to purified antibodies, our production only takes 2 weeks — and with RushData, our AI antibody wet lab validation service, computational predictions become experimentally confirmed candidates at high throughput. We have delivered tens of thousands of recombinant antibodies for more than 3,000 biotech and pharma companies worldwide.

‍